This project focuses on building an object detection model to detect bark beetles — a small insects that threaten forests if invasive. The aim is for the model to detect both individual beetles and clusters of beetles, whether in the wild or in lab petri dishes. The model will be used as part of a image classifier model to identify the species.
General-purpose detection models tend to underperform on small-scale objects like bark beetles, especially when:
- Beetles are very small in the image.
- Beetles are densely clustered or partially occluded.
- Beetles appear in varying backgrounds — from bark textures to sterile lab environments.
Our objective is to create a dataset and a model that can robustly detect bark beetles across diverse image settings.
Initial Dataset Size: ~58,000 images
Challenge: No bounding box annotations available.
To quick start the annotation process, I developed a semi-supervised pipeline using GroundingDINO to generate initial bounding boxes. These labels were then used to train a series of YOLOv10 models in iterative fashion. Meaning, the next model will train on the labels generated from the previous model.
The number of detections per iteration was used as a proxy metric for model improvement, since ground truth annotations were not available yet.
| Model | Detections |
|---|---|
| GroundingDINO | 173,503 |
| YOLOv10 - v1 | 177,127 |
| YOLOv10 - v2 | 178,396 |
| YOLOv10 - v3 | 179,681 |
| YOLOv10 - v4 | 180,557 |
| YOLOv10 - v5 | 210,098 |
| YOLOv10 - v6 | 211,986 |
| YOLOv10 - v7 | 215,083 |
| YOLOv10 - v8 | 217,753 |
| YOLOv10 - v9 | 219,882 |
We can see that with each iteration, the number of detections steadily increased. This created a strong starting point for human-in-the-loop label correction, allowing manual reviewers to refine the bounding boxes.
To make the label refinement process more efficient, I developed a bounding box editor inside a Jupyter Notebook. This tool allows users to:
- Quickly view each image with its predicted bounding boxes.
- Delete incorrect boxes (e.g., boxes around rocks, overlapping beetles, or multiple beetles in one box).
- Navigate efficiently through the dataset.
In addition to removing bad labels, I also observed missing detections — where beetles are present but not labeled. To handle this, the editor was used in conjunction with the labelme annotation tool to manually add missing bounding boxes.
Thanks to the reduced editing required from the semi-supervised labeling pipeline, we were able to efficiently create a fully labeled object detection dataset of 58,000 images.
With the finalized labeled dataset, I am now working on benchmarking a set of state-of-the-art object detection architectures:
- YOLOv10
- YOLOv11
- DETR (DEtection TRansformer)
- Co-DETR (Context-enhanced DETR)
- RF-DETR
To ensure robust evaluation, I am using 5-fold cross-validation to compare each model's performance across metrics such as:
- Mean Average Precision (mAP)
- Precision / Recall
- Inference Speed
- Detection robustness on small clustered objects
The best-performing model will be selected
| Fold | YOLOv10 | Co-DETR |
|---|---|---|
| 1 | 0.97848 | 0.931 |
| 2 | 0.95066 | 0.923 |
| 3 | 0.97371 | 0.942 |
| 4 | 0.96982 | 0.930 |
| 5 | 0.97123 | 0.945 |
More results soon to come!